diff --git a/build_use.sh b/build_use.sh
index e9e0d3b..16c9aac 100644
--- a/build_use.sh
+++ b/build_use.sh
@@ -8,9 +8,9 @@ cp meta/meta.json use_model/meta.json
 # create the package
 mkdir -p use_package
 python -m spacy package use_model use_package --force
-pushd use_package/en_use-0.1.2
+pushd use_package/en_use-0.1.3
 # zip it
 python setup.py sdist
 # install the tar.gz from dist/en_use-0.1.1.tar.gz
-pip install dist/en_use-0.1.2.tar.gz
+pip install dist/en_use-0.1.3.tar.gz
 popd
\ No newline at end of file
diff --git a/meta/meta.json b/meta/meta.json
index 2d10ba0..2472169 100644
--- a/meta/meta.json
+++ b/meta/meta.json
@@ -1,7 +1,7 @@
 {
     "lang": "en",
     "name": "use",
-    "version": "0.1.2",
+    "version": "0.1.3",
     "spacy_version": ">=2.2.3",
     "description": "Using TFHub USE",
     "author": "Martino Mensio",
diff --git a/setup.py b/setup.py
index 3d05841..06d7a4d 100644
--- a/setup.py
+++ b/setup.py
@@ -1,8 +1,10 @@
 from setuptools import setup, find_packages
-# to import the version and also run the module one time (download cache model)
+# to import the version
 import universal_sentence_encoder
 
 def setup_package():
+    # run the module one time (download cache model)
+    universal_sentence_encoder.UniversalSentenceEncoder.create_wrapper()
     setup(
         name="universal_sentence_encoder",
         entry_points={
diff --git a/universal_sentence_encoder/__init__.py b/universal_sentence_encoder/__init__.py
index 2becc76..4b49d82 100644
--- a/universal_sentence_encoder/__init__.py
+++ b/universal_sentence_encoder/__init__.py
@@ -9,7 +9,7 @@
 from spacy.tokens import Span
 from spacy.matcher import Matcher
 
-__version__ = "0.1.0"
+__version__ = "0.1.3"
 
 from .language import UniversalSentenceEncoder
 UniversalSentenceEncoder.install_extensions()
@@ -30,7 +30,8 @@ def __init__(self, nlp, enable_cache):
         # enable_cache = cfg.get('enable_cache', True)
         # UniversalSentenceEncoder.install_extensions()
         print('enable_cache', enable_cache)
-        UniversalSentenceEncoder.tf_wrapper.enable_cache = enable_cache
+        # load tfhub now (not compulsory but nice to have it loaded when running `spacy.load`)
+        UniversalSentenceEncoder.create_wrapper(enable_cache=enable_cache)
 
     def __call__(self, doc):
         UniversalSentenceEncoder.overwrite_vectors(doc)
diff --git a/universal_sentence_encoder/language.py b/universal_sentence_encoder/language.py
index 0afd8d4..f1874b6 100644
--- a/universal_sentence_encoder/language.py
+++ b/universal_sentence_encoder/language.py
@@ -16,10 +16,22 @@ class UniversalSentenceEncoder(Language):
 
     @staticmethod
     def install_extensions():
-        UniversalSentenceEncoder.tf_wrapper = TFHubWrapper()
+        def get_encoding(token_span_doc):
+            # tokens, spans and docs all have the `.doc` property
+            wrapper = token_span_doc.doc._.tfhub_wrapper
+            if wrapper == None:
+                raise ValueError('Wrapper None')
+            return wrapper.embed_one(token_span_doc)
+        
+        # Placeholder for a reference to the wrapper
+        Doc.set_extension('tfhub_wrapper', default=None, force=True)
         # set the extension both on doc and span level
-        Span.set_extension('universal_sentence_encoding', getter=UniversalSentenceEncoder.tf_wrapper.embed_one, force=True)
-        Doc.set_extension('universal_sentence_encoding', getter=UniversalSentenceEncoder.tf_wrapper.embed_one, force=True)
+        # Token.set_extension('universal_sentence_encoding', getter=UniversalSentenceEncoder.tf_wrapper.embed_one, force=True)
+        # Span.set_extension('universal_sentence_encoding', getter=UniversalSentenceEncoder.tf_wrapper.embed_one, force=True)
+        # Doc.set_extension('universal_sentence_encoding', getter=UniversalSentenceEncoder.tf_wrapper.embed_one, force=True)
+        Token.set_extension('universal_sentence_encoding', getter=get_encoding, force=True)
+        Span.set_extension('universal_sentence_encoding', getter=get_encoding, force=True)
+        Doc.set_extension('universal_sentence_encoding', getter=get_encoding, force=True)
 
     @staticmethod
     def overwrite_vectors(doc):
@@ -33,6 +45,9 @@ def overwrite_vectors(doc):
         # doc.user_hooks["vector_norm"] = lambda a: a._.universal_sentence_encoding
         # doc.user_span_hooks["vector_norm"] = lambda a: a._.universal_sentence_encoding
         # doc.user_token_hooks["vector_norm"] = lambda a: a._.universal_sentence_encoding
+        
+        # save a reference to the wrapper
+        doc._.tfhub_wrapper = TFHubWrapper.get_instance()
         return doc
 
 
@@ -44,6 +59,17 @@ def create_nlp(language_base='en'):
         nlp.add_pipe(UniversalSentenceEncoder.overwrite_vectors)
         return nlp
 
+    # def __init__(self, vocab=True, make_doc=True, max_length=10 ** 6, meta={}, **kwargs):
+    #     self.tf_wrapper = TFHubWrapper.get_instance()
+    #     super.__init__(self, vocab, make_doc, max_length, meta=meta, **kwargs)
+
+    @staticmethod
+    def create_wrapper(enable_cache=True):
+        """Helper method, run to do the loading now"""
+        UniversalSentenceEncoder.tf_wrapper = TFHubWrapper.get_instance()
+        # TODO the enable_cache with singleton is not a great idea
+        UniversalSentenceEncoder.tf_wrapper.enable_cache = enable_cache
+
 class UniversalSentenceEncoderPipe(Pipe):
     pass
 
@@ -51,6 +77,15 @@ class UniversalSentenceEncoderPipe(Pipe):
 class TFHubWrapper(object):
     embed_cache: Dict[str, Any]
     enable_cache = True
+    instance = None
+
+    @staticmethod
+    def get_instance():
+        # singleton
+        if not TFHubWrapper.instance:
+            TFHubWrapper.instance = TFHubWrapper()
+        return TFHubWrapper.instance
+
 
     def __init__(self):
         self.embed_cache = {}